--- license: apache-2.0 datasets: - cerebras/SlimPajama-627B - bigcode/starcoderdata - mc4 language: - fi - en - da - sv - 'no' - nn - is ---
# Viking 33B Model Card _**NOTE:** This is a **research checkpoint** of a model for which **training has not been completed.** It is being provided in its current state for research and testing purposes. **Care should be taken when using the outputs of the model.** Once pretraining has completed we intend to release additional instruction-tuned and chat-tuned varieties._ Viking 33B is a 33B parameter decoder-only transformer pretrained on Finnish, English, Swedish, Danish, Norwegian, Icelandic and code. It is being trained on 2 trillion tokens (700B billion as of this release). Viking 33B is a fully open source model and is made available under the Apache 2.0 License. Viking was created in a collaboration between [SiloGen](https://www.silo.ai/silogen) from [Silo AI](https://www.silo.ai/), the [TurkuNLP group](https://turkunlp.org/) of the University of Turku, and [High Performance Language Technologies](https://hplt-project.org/) (HPLT). Training was conducted on the [LUMI supercomputer](https://www.lumi-supercomputer.eu/), using compute resources generously provided by [CSC](https://csc.fi/) - IT Center for Science, Finland. This project is part of an ongoing effort to create open source large language models for non-English and especially low resource languages like Finnish. The mode is fluent in Finnish, English, the Scandinavian languages and capable of basic translation between them. It is also able to understand and generate code. Viking 33B is part of the second set of models of our model family. Work is also underway on our next models which will support even more languages and grouped query attention. ## Model Overview _**NOTE:** In addition to being an early research release, Viking is a base model which needs further fine tuning for most use cases._ Viking is a generative pretrained transformer using a LLaMA-like GPT architecture, and makes use of rotary positional embeddings and flash attention. | Hyperparameter | Value | | :------------- | :----: | | n_parameters | 33B | | n_layers | 56 | | n_heads | 56 | | d_model | 7168 | | vocab_size | 131072 | | sequence_length | 4096 | ## Viking Research Checkpoints Checkpoints are available as branches in the repository. Checkpoints will be released roughly every 100B tokens. The main branch will always point to the latest checkpoint. The following checkpoints are available: * [100B](https://huggingface.co/LumiOpen/Viking-33B/tree/100B) * [200B](https://huggingface.co/LumiOpen/Viking-33B/tree/200B) * [300B](https://huggingface.co/LumiOpen/Viking-33B/tree/300B) * [400B](https://huggingface.co/LumiOpen/Viking-33B/tree/400B) * [500B](https://huggingface.co/LumiOpen/Viking-33B/tree/500B) * [600B](https://huggingface.co/LumiOpen/Viking-33B/tree/600B) * [700B](https://huggingface.co/LumiOpen/Viking-33B/tree/700B) * [800B](https://huggingface.co/LumiOpen/Viking-33B/tree/800B) * [900B](https://huggingface.co/LumiOpen/Viking-33B/tree/900B) * [1000B](https://huggingface.co/LumiOpen/Viking-33B/tree/1000B) The transformers library allows you to load a checkpoint from a branch as follows: ```python branch = "200B" model = transformers.AutoModelForCausalLM.from_pretrained( "LumiOpen/Viking-33B", torch_dtype=torch.bfloat16, revision=branch, ) ``` ## Training Viking 33B was trained on the LUMI supercomputer, using 1024 AMD MI250X GPUs. Each MI250X GPU has two Graphics Complex Dies (GCDs) for a world size of 2048 during training, using activation checkpointing, a micro batch size of 1, gradient accumulation of 16, and a 3D parallelism strategy of TP=4, PP=4, DP=128. Training began in September 2023 using a [custom fork](https://github.com/LumiOpen/Megatron-DeepSpeed) of the Megatron-Deepspeed framework. ## Training Hyperparameters | Hyperparameter | Value | Comment | | :------------: | :---: | :------:| | Precision | bfloat16 | | | Optimizer | AdamW | | | Learning rate | 3e-4 | 10B tokens warm-up, cosine decay to 3e-5 | | Weight decay | 1e-1 | | | Batch size | 1024 | 1024 samples x 4096 tokens = 4194304 tokens | ## Tokenizer Viking uses a custom 128K Bloom tokenizer trained on the same English, Finnish, Swedish, Danish, Norwegian, Icelandic and code dataset used to train the model. ## Dataset Viking is being trained on a 2 trillion token mixed dataset of English, Finnish, Swedish, Danish, Norwegian, Icelandic and code. Full details will be published soon. ## Evaluation Results Full evaluation results will be published with the final model. ## Ethical Considerations and Limitations _Viking 33B is a release of a partially trained model, and special care should be taken when using any output._ Viking is an advanced language model, primarily optimized for English, Finnish, Swedish, Norwegian, Danish, Icelandic and code, with no meaningful proficiency in any other languages. As with most AI-driven systems, Viking is a product of the vast data it has been trained on, which may reflect the imperfections, biases, and idiosyncrasies of the wider web. Viking may, at times, produce outputs that can be considered inaccurate, prejudiced, or controversial. Users and developers engaging with Viking should exercise discretion and consider additional evaluation and customization to ensure the model's responses align with their specific needs and ethical standards. ## License Viking is released under the Apache 2.0 license.